feat: Vector Search microservice + parallel RAG engine

[ShaerWare]

Summary

• Add standalone Vector Search microservice (services/vector-search/) — FastAPI + ChromaDB + paraphrase-multilingual-mpnet-base-v2 (768 dims), Bearer token auth, persistent storage
• Add async HTTP client (app/services/vector_search_client.py) — httpx with retry + graceful degradation
• Integrate as parallel search engine in WikiRAGService: new search_async(), retrieve_async(), retrieve_multi_async() run BM25 + embeddings + vector search via asyncio.gather, merge/deduplicate results
• Switch all 4 RAG injection call sites in chat router to async (_inject_rag_context_async)
• Add background sync task (vector-search-sync) + incremental sync via DatasetSynced events
• Add admin endpoints: GET /admin/wiki-rag/vector-search/status, POST .../sync
• Add vector search status to /health endpoint
• Docker: vector-search service (profile: vector-search, port 8003), vector_search_data volume
• Env vars: VECTOR_SEARCH_URL, VECTOR_SEARCH_TOKEN
• Update CLAUDE.md with new architecture docs

Relates to #489

NEWS

🔍 Vector Search — новый движок семантического поиска

Добавлен микросервис Vector Search на базе ChromaDB и мультиязычной модели paraphrase-multilingual-mpnet-base-v2. Работает параллельно с BM25 и embeddings — все три движка ищут одновременно, результаты объединяются и дедуплицируются. Запускается через Docker: docker compose --profile vector-search up -d.

Test plan

• docker compose --profile vector-search up -d — microservice starts
• curl http://localhost:8003/health — returns ok + model info
• Restart orchestrator with VECTOR_SEARCH_URL=http://localhost:8003
• GET /admin/wiki-rag/vector-search/status — shows connected
• POST /admin/wiki-rag/vector-search/sync — upserts all sections
• POST /admin/wiki-rag/search — results include engine: "vector_search"
• Chat with RAG enabled — vector search results included in context
• ruff check . + cd admin && npm run lint:check pass
• All 106 unit tests pass

🤖 Generated with Claude Code